In real life, visual learning is supposed to be a continuous process. Opposite to this, in machine vision, learning has been so far considered in a limited and isolated way. Most of the nowadays approaches, with very few exceptions, require the intervention of the human operator to collect, segment and store hand-picked data and train pattern classifiers with them. However, for real world scenarios, it is unlikely to know beforehand the number of total classes or the exact number of instances per class.
There is a recent trend in pattern recognition represented by online learning approaches, which treat to continuously update the data representation when new information arrives. Starting with a minimal set, the initial knowledge is expanded, by incorporating incoming instances, some of them which have not been previously available or foreseen at the application creation stage. An interesting characteristic of this strategy is that the train and test phases take place simultaneously.
Given the increasing interest on this subject, the current call for papers is intended to offer the possibility to discuss the most recent advances in the area.
This special session is intended to cover, but it is not limited to, the following topics:
- Feature extraction and representation
- Incremental/decremental learning strategies
- Online classification
- Unsupervised clustering
- Novelty detection
- Knowledge discovery
- Optimal data representation to manage large datasets
- Extraction and creation of semantic content from data
- Contextual representation and reasoning
- Developmental architectures
- Modelling of cognitive processes